IBM watsonx

IBM watsonx is IBM’s enterprise generative AI and data platform. The directory focus is watsonx.ai: the studio for prompting, RAG, agents (AgentLab), fine-tuning, and inferencing IBM Granite plus third-party models (Meta, Mistral, DeepSeek, and others) on IBM Cloud / hybrid. Sister products watsonx.data and watsonx.governance complete the suite but are separate SKUs. As of September 2026 IBM publishes plan tiers and indicative token rates on ibm.com/products/watsonx-ai/pricing and docs.

Verified plan framing: Free / Lite playground with up to 300,000 tokens/month for foundation model inferencing, 20 CUH ML tools, 100 text-extraction pages; Essentials pay-as-you-go from $0/month platform fee plus metered usage; Standard enterprise production starting at $1,110/month including a block of capacity (docs: 2,500 CUH) with overage and tokens metered; HIPAA-Ready instance fee $1,800/month at Standard usage rates (docs). Example Granite rate on the pricing page: granite-4h-small USD 0.0636 / 1M input and 0.265 / 1M output. Prices are indicative and may vary by country: verify pricing in the catalog at purchase.

What watsonx.ai delivers as of September 2026

PromptLab for foundation-model work, RAG patterns, AgentLab for agents, synthetic data generator, ML tooling, text extraction, LoRA/QLoRA fine-tuning (plan-gated), custom foundation models and deploy-on-demand hosting on higher plans. Model library spans IBM Granite 4.x / 3.x families (including granite-4-h-small, vision, speech, code, guardian) and third-party instruct models. Inference billing uses Resource Units (1 RU ≈ 1,000 tokens) on runtime plans. Regional model availability differs by data center.

Where watsonx sits versus Bedrock and Azure OpenAI

Amazon Bedrock and Azure OpenAI are hyperscaler model API gateways with strong ecosystem lock-in to AWS/Azure. watsonx competes for IBM Cloud / Red Hat / regulated enterprises that want Granite plus governance narrative (watsonx.governance) and hybrid deployment. ChatGPT Enterprise is the assistant UX peer; watsonx.ai is a builder studio, not a consumer chat app.

"Enterprise production: Starting at USD 1110/month" - IBM, ibm.com/products/watsonx-ai/pricing, retrieved September 2026

Cost and friction

Lite is evaluation-scale (300k tokens/month). Essentials avoids the $1,110 floor but can surprise on CUH + RU + page extraction. Standard’s $1,110 includes CUH block but tokens and extra CUH still meter. HIPAA-Ready adds $1,800 instance fee. Advanced support SLAs start at $200/month on the pricing footnote. Cross-border list prices vary: IBM’s own disclaimer says indicative.

Where it still breaks

Not a polished consumer assistant. Catalog HTML often omits every Granite cell’s dollar figure (some rows render blank in page extracts): always confirm the live model table or Cloud catalog SKU. Tuning experiments are unavailable on Lite. Teams outside IBM Cloud gravity may prefer Bedrock/Vertex for simpler token shopping.

"For foundation model inference, charges are based on a Resource Unit (RU) metric equivalent to 1000 tokens (including both input and output tokens)." - IBM watsonx.ai pricing footnotes, retrieved September 2026

Who should use watsonx: and who should skip it

Use watsonx.ai for enterprise builders on IBM Cloud who need Granite, hybrid options, and governance adjacent to watsonx.governance. Start on Free/Lite, then Essentials for PAYG, Standard when you need the enterprise capacity floor. Skip if you want ChatGPT-like UX only, or a single cheap frontier API without IBM platform fees.